The image data of upper body from a depth sensor is used to estimate the state of human focusing on the incidents that might happen while using a walker. Several falling cases along with sitting and normal walking are considered in this study. Two main features namely the centroid and the principal component analysis (PCA) values of the upper body are used to classify the data. The non-walking states are detected either by using a Gaussian Mixture Model of PCA features or training a Continuous Hidden Markov Model (CHMM) with centroid data. The CHMM is also used to detect the type of falling. The state estimation results are used to control the motion of a passive type walker referred to as RT Walker. Falling prevention and sitting/standing assistance are achieved using both methods. Performance of the methods are discussed and compared to each other from different aspect.